/run-ai-task

Print a ready-to-run command that launches the containerized AI coding task.

The runner spawns a Docker container in which Claude Code works on the target Flutter project interactively. If this skill executes the runner itself, its stdout/stdin get hidden behind the slash-command turn and the user cannot watch or interact with the inner Claude session. So this skill MUST stop at "print the command" and let the user run it from a separate terminal.

Arguments (positional): $ARGUMENTS

1. app_dir — path to the target Flutter app (must be a clean git working tree)
2. prompt_file — path to the task prompt (.txt/.md)
3. task_name — short slug used in the run id
4. setup_file (optional) — a bootstrap script run inside the container before flutter pub get (e.g. melos bs, make init, code generation)

Steps:

1. Validate that the first three arguments are present. If not, ask the user for the missing ones. setup_file is optional — omit it for a plain flutter pub get.

2. Print the command for the user to copy-paste into a separate terminal. Use the form:

   "${CLAUDE_PLUGIN_ROOT}/run-ai-task.sh" <app_dir> <prompt_file> <task_name> [setup_file]
   

   Resolve ${CLAUDE_PLUGIN_ROOT} to its actual absolute path in the printed command so the user can paste it verbatim. Quote any path that contains spaces.

3. Tell the user to run it in a separate terminal so they can watch the inner Claude Code session and interact with it if needed.

4. Briefly note what to expect after it finishes: a run directory under runs/<RUN_ID>/ and an exported patch at runs/<RUN_ID>/output/ai-change.patch, plus the git apply --check / git apply --stat hints that the script prints for reviewing the change before applying it.

Do NOT invoke run-ai-task.sh from this skill — only print the command. Do NOT apply the resulting patch automatically — exporting it for human review is the whole point.